Evaluation of Sentinel-1 SAR Interferometry to Improve Forest Biomass Estimation from Multisource Satellite Data, Field Plots and Other Data in Thailand
Time-series coherence maps, derived from Sentinel-1 C-band SAR images used in the AGB model, are analysed in this study. Several factors impacted the coherence losses due to the temporal decorrelation, perpendicular baseline, weather conditions and the source of DEM. These factors were related to the accuracy of the AGB model. Consequently, the analysis factors with different pair images to extract the coherence map, the AGB model illustrated an RMSE range from 66.8 t∙ha-1 to 89.4 t∙ha-1. Therefore, by selecting suitable pair images of Sentinel-1 and configuring the control parameters in machine learning, it is possible to improve the accuracy of the AGB map. In addition, the fire frequency map has the potential to enhance the AGB model when combined with the coherence map acquired during the rainy season (under the condition of the average rainfall intensity <30 mm/day and wind speed <7.5 km/h).
In the comparison between two machine learning approaches (maximum entropy [MaxEnt] and random forest [RF]), the MaxEnt represented R2 = 0.58 and RMSE = 55.8 t∙ha-1 better than the AGB generated from the RF model (R2 = 0.42 and RMSE = 67.4 t∙ha-1). However, when compared at each biomass class interval, the MaxEnt showed a high performance of AGB>120 t/ha-1. This was in contrast with the RF model, which was better in AGB <120 t∙ha-1. Further, the MaxEnt algorithm had the potential to transfer into the other area (unavailable forest plots) using six variable factors. Furthermore, it obtained similar accuracy to the original map. Meanwhile, the use of more than six variable factors tended to obtain low accuracy AGB model.
The investigation of different spatial patterns of the AGB map by using the fuzzy numerical index (FNI) was compared between the GlobBiomass map (global) and the MaxEnt model map (local). The results highlighted that the AGB in rubber plantation and deciduous forest areas were dissimilar between the two maps. Therefore, the development of a specific AGB model for each forest type could be improved accuracy and reduced an error in the tropical forest zone.
Supervisor(s)Heiko Balzter, Susan Page
Date of award2023-01-27
Author affiliationDepartment of Geography
Awarding institutionUniversity of Leicester